NW: AspectC++ support

8 files changed, 87 insertions, 58 deletions

diff --git a/NW/Makefile b/NW/Makefile
index 68f495a..30353c9 100644
--- a/NW/Makefile
+++ b/NW/Makefile
@@ -1,46 +1,56 @@
-DPU_DIR := dpu
-HOST_DIR := host
-BUILDDIR ?= bin
 NR_TASKLETS ?= 13
 BL ?= 1024 
 BL_IN ?= 4 
 NR_DPUS ?= 1 
 ENERGY ?= 0
 
-define conf_filename
-	${BUILDDIR}/.NR_DPUS_$(1)_NR_TASKLETS_$(2)_BL_$(3).conf
-endef
-CONF := $(call conf_filename,${NR_DPUS},${NR_TASKLETS},${BL})
+HOST_SOURCES := $(wildcard host/*.c)
+DPU_SOURCES := $(wildcard dpu/*.c)
 
-HOST_TARGET := ${BUILDDIR}/nw_host
-DPU_TARGET := ${BUILDDIR}/nw_dpu
+aspectc ?= 0
+aspectc_timing ?= 0
 
-COMMON_INCLUDES := support
-HOST_SOURCES := $(wildcard ${HOST_DIR}/*.c)
-DPU_SOURCES := $(wildcard ${DPU_DIR}/*.c)
+HOST_CC := ${CC}
 
-.PHONY: all clean test
+COMMON_FLAGS := -Wall -Wextra -g -Iinclude -DNR_TASKLETS=${NR_TASKLETS} -DNR_DPUS=${NR_DPUS} -DBL=${BL}
+HOST_FLAGS := ${COMMON_FLAGS} -O3 `dpu-pkg-config --cflags --libs dpu` -DENERGY=${ENERGY} -DASPECTC=${aspectc}
+DPU_FLAGS := ${COMMON_FLAGS} -O2 -DBL_IN=${BL_IN}
+
+ifeq (${aspectc_timing}, 1)
+	ASPECTC_HOST_FLAGS += -ainclude/dfatool_host_dpu.ah -ainclude/dfatool_host.ah
+endif
+
+ASPECTC_HOST_FLAGS ?= -a0
 
-__dirs := $(shell mkdir -p ${BUILDDIR})
+ifeq (${aspectc}, 1)
+	HOST_CC = ag++ -r repo.acp -v 0 ${ASPECTC_HOST_FLAGS} --c_compiler ${UPMEM_HOME}/bin/clang++ -p . --Xcompiler
+else
+	HOST_FLAGS += -std=c11
+endif
 
-COMMON_FLAGS := -Wall -Wextra -g -I${COMMON_INCLUDES}
-HOST_FLAGS := ${COMMON_FLAGS} -std=c11 -O3 `dpu-pkg-config --cflags --libs dpu` -DNR_TASKLETS=${NR_TASKLETS} -DNR_DPUS=${NR_DPUS} -DBL=${BL} -DENERGY=${ENERGY}
-DPU_FLAGS := ${COMMON_FLAGS} -O2 -DNR_TASKLETS=${NR_TASKLETS} -DBL=${BL} -DBL_IN=${BL_IN}
+QUIET = @
 
-all: ${HOST_TARGET} ${DPU_TARGET}
+ifdef verbose
+	QUIET =
+endif
 
-${CONF}:
-	$(RM) $(call conf_filename,*,*)
-	touch ${CONF}
+all: bin/nw_host bin/nw_dpu
 
-${HOST_TARGET}: ${HOST_SOURCES} ${COMMON_INCLUDES} ${CONF}
-	$(CC) -o $@ ${HOST_SOURCES} ${HOST_FLAGS}
+bin:
+	${QUIET}mkdir -p bin
 
-${DPU_TARGET}: ${DPU_SOURCES} ${COMMON_INCLUDES} ${CONF}
-	dpu-upmem-dpurte-clang ${DPU_FLAGS} -o $@ ${DPU_SOURCES}
+bin/nw_host: ${HOST_SOURCES} include bin
+	${QUIET}cp ../include/dfatool_host_dpu.ah include
+	${QUIET}${HOST_CC} -o $@ ${HOST_SOURCES} ${HOST_FLAGS}
+	${QUIET}rm -f include/dfatool_host_dpu.ah
+
+bin/nw_dpu: ${DPU_SOURCES} include bin
+	${QUIET}dpu-upmem-dpurte-clang ${DPU_FLAGS} -o $@ ${DPU_SOURCES}
 
 clean:
 	$(RM) -r $(BUILDDIR)
 
 test: all
-	./${HOST_TARGET}
+	bin/nw_host
+
+.PHONY: all clean test
diff --git a/NW/dpu/task.c b/NW/dpu/task.c
index c022f70..fab163a 100644
--- a/NW/dpu/task.c
+++ b/NW/dpu/task.c
@@ -10,7 +10,7 @@
 #include <perfcounter.h>
 #include <barrier.h>
 
-#include "../support/common.h"
+#include "common.h"
 
 __host dpu_arguments_t DPU_INPUT_ARGUMENTS;
 
diff --git a/NW/host/app.c b/NW/host/app.c
index 0e899ec..9de2918 100644
--- a/NW/host/app.c
+++ b/NW/host/app.c
@@ -7,20 +7,30 @@
 #include <stdlib.h>
 #include <stdbool.h>
 #include <string.h>
+
+#if ASPECTC
+extern "C" {
+#endif
+
 #include <dpu.h>
 #include <dpu_log.h>
-#include <unistd.h>
-#include <getopt.h>
-#include <assert.h>
-
-#include "../support/common.h"
-#include "../support/timer.h"
-#include "../support/params.h"
 
 #if ENERGY
 #include <dpu_probe.h>
 #endif
 
+#if ASPECTC
+}
+#endif
+
+#include <unistd.h>
+#include <getopt.h>
+#include <assert.h>
+
+#include "common.h"
+#include "timer.h"
+#include "params.h"
+
 // Define the DPU Binary path as DPU_BINARY here
 #ifndef DPU_BINARY
 #define DPU_BINARY "./bin/nw_dpu"
@@ -184,7 +194,7 @@ int main(int argc, char **argv) {
 
     struct Params p = input_params(argc, argv);
     struct dpu_set_t dpu_set, dpu;
-    uint32_t nr_of_dpus, max_dpus;
+    uint32_t nr_of_dpus, nr_of_ranks, max_dpus;
 
 #if ENERGY
     struct dpu_probe_t probe;
@@ -195,6 +205,7 @@ int main(int argc, char **argv) {
     DPU_ASSERT(dpu_alloc(NR_DPUS, NULL, &dpu_set));
     DPU_ASSERT(dpu_load(dpu_set, DPU_BINARY, NULL));
     DPU_ASSERT(dpu_get_nr_dpus(dpu_set, &nr_of_dpus));
+    DPU_ASSERT(dpu_get_nr_ranks(dpu_set, &nr_of_ranks));
     printf("Allocated %d DPU(s)\n", nr_of_dpus);
     printf("Allocated %d TASKLET(s) per DPU\n", NR_TASKLETS);
 #if DYNAMIC
@@ -822,28 +833,6 @@ int main(int argc, char **argv) {
             stop(&timer, 1);
 
     }
-
-    // Print timing results
-    printf("CPU version ");
-    print(&timer, 0, p.n_reps);
-    printf("CPU-DPU ");
-    print(&timer, 2, p.n_reps);
-    printf("DPU Kernel ");
-    print(&timer, 3, p.n_reps);
-    printf("Inter-DPU ");
-    print(&timer, 1, p.n_reps);
-    printf("DPU-CPU ");
-    print(&timer, 4, p.n_reps);
-    printf("\n");
-    printf("Longest Diagonal CPU-DPU ");
-    print(&long_diagonal_timer, 2, p.n_reps);
-    printf("Longest Diagonal DPU Kernel ");
-    print(&long_diagonal_timer, 3, p.n_reps);
-    printf("Longest Diagonal Inter-DPU ");
-    print(&long_diagonal_timer, 1, p.n_reps);
-    printf("Longest Diagonal DPU-CPU ");
-    print(&long_diagonal_timer, 4, p.n_reps);
-    printf("\n");
     
 #if ENERGY
     printf("DPU Energy (J): %f \t ", tavg_energy / p.n_reps);
diff --git a/NW/support/common.h b/NW/include/common.h
index 69069e7..69069e7 100755..100644
--- a/NW/support/common.h
+++ b/NW/include/common.h
diff --git a/NW/include/dfatool_host.ah b/NW/include/dfatool_host.ah
new file mode 100644
index 0000000..d45aef3
--- /dev/null
+++ b/NW/include/dfatool_host.ah
@@ -0,0 +1,30 @@
+#pragma once
+
+#include <sys/time.h>
+#include "dfatool_host_dpu.ah"
+
+aspect DfatoolHostTiming : public DfatoolHostDPUTiming {
+
+	unsigned long n_elements;
+	unsigned int element_size;
+
+	virtual int getKernel() { return 1; }
+
+	DfatoolHostTiming() {
+		element_size = sizeof(uint32_t);
+	}
+
+	advice call("% input_params(...)"): after() {
+		Params* p = tjp->result();
+		n_elements = p->max_rows;
+		printf("[>>] NW | n_dpus=%u n_elements=%lu\n", NR_DPUS, n_elements);
+	}
+
+	advice call("% srand(...)") : after() {
+		printf("[--] NW | n_dpus=%u n_elements=%lu\n", NR_DPUS, n_elements);
+	}
+
+	advice execution("% main(...)") : after() {
+		printf("[<<] NW | n_dpus=%u n_elements=%lu\n", NR_DPUS, n_elements);
+	}
+};
diff --git a/NW/support/params.h b/NW/include/params.h
index 8874248..8874248 100644
--- a/NW/support/params.h
+++ b/NW/include/params.h
diff --git a/NW/support/timer.h b/NW/include/timer.h
index efaefcd..efaefcd 100755..100644
--- a/NW/support/timer.h
+++ b/NW/include/timer.h
diff --git a/README.md b/README.md
index b77dbf3..307008e 100644
--- a/README.md
+++ b/README.md
@@ -77,7 +77,7 @@ CPU and DPU benchmarks in this repository have been adjusted as follows:
 * HST-L: AD
 * HST-S: DLN
 * MLP: A
-* NW: –
+* NW: A
 * RED: DLN
 * SCAN-SSA: D
 * SCAN-RSS: DLN